Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Standardization of Behavioral Use Clauses and Their Adoption for Responsible Licensing of AI

Daniel McDuff, Tim Korjakow, Scott Cambo, Jesse Josua Benjamin, Jenny Lee, Yacine Jernite, Carlos Muñoz Ferrandis, Aaron Gokaslan, Alek Tarkowski, Joseph Lindley, A. Feder Cooper, Danish Contractor

TL;DR

The paper investigates why behavioral-use clauses in AI licenses have proliferated and how they are adopted, using a mixed-methods analysis of over 170k repositories, clause clustering, and semi-structured interviews. It finds rapid growth and substantial variation in RAIL licenses, argues that standardization is necessary to prevent user confusion, and demonstrates that domain-specific customization remains valuable in some contexts. The authors propose a practical path forward: standardized customization enabled by tooling, including a license generator and dependency-scanning support, to balance responsible use with openness. This work has practical implications for governance and open-source AI, offering concrete mechanisms to scale responsible licensing across diverse AI assets and ecosystems.

Abstract

Growing concerns over negligent or malicious uses of AI have increased the appetite for tools that help manage the risks of the technology. In 2018, licenses with behaviorial-use clauses (commonly referred to as Responsible AI Licenses) were proposed to give developers a framework for releasing AI assets while specifying their users to mitigate negative applications. As of the end of 2023, on the order of 40,000 software and model repositories have adopted responsible AI licenses licenses. Notable models licensed with behavioral use clauses include BLOOM (language) and LLaMA2 (language), Stable Diffusion (image), and GRID (robotics). This paper explores why and how these licenses have been adopted, and why and how they have been adapted to fit particular use cases. We use a mixed-methods methodology of qualitative interviews, clustering of license clauses, and quantitative analysis of license adoption. Based on this evidence we take the position that responsible AI licenses need standardization to avoid confusing users or diluting their impact. At the same time, customization of behavioral restrictions is also appropriate in some contexts (e.g., medical domains). We advocate for ``standardized customization'' that can meet users' needs and can be supported via tooling.

On the Standardization of Behavioral Use Clauses and Their Adoption for Responsible Licensing of AI

TL;DR

The paper investigates why behavioral-use clauses in AI licenses have proliferated and how they are adopted, using a mixed-methods analysis of over 170k repositories, clause clustering, and semi-structured interviews. It finds rapid growth and substantial variation in RAIL licenses, argues that standardization is necessary to prevent user confusion, and demonstrates that domain-specific customization remains valuable in some contexts. The authors propose a practical path forward: standardized customization enabled by tooling, including a license generator and dependency-scanning support, to balance responsible use with openness. This work has practical implications for governance and open-source AI, offering concrete mechanisms to scale responsible licensing across diverse AI assets and ecosystems.

Abstract

Growing concerns over negligent or malicious uses of AI have increased the appetite for tools that help manage the risks of the technology. In 2018, licenses with behaviorial-use clauses (commonly referred to as Responsible AI Licenses) were proposed to give developers a framework for releasing AI assets while specifying their users to mitigate negative applications. As of the end of 2023, on the order of 40,000 software and model repositories have adopted responsible AI licenses licenses. Notable models licensed with behavioral use clauses include BLOOM (language) and LLaMA2 (language), Stable Diffusion (image), and GRID (robotics). This paper explores why and how these licenses have been adopted, and why and how they have been adapted to fit particular use cases. We use a mixed-methods methodology of qualitative interviews, clustering of license clauses, and quantitative analysis of license adoption. Based on this evidence we take the position that responsible AI licenses need standardization to avoid confusing users or diluting their impact. At the same time, customization of behavioral restrictions is also appropriate in some contexts (e.g., medical domains). We advocate for ``standardized customization'' that can meet users' needs and can be supported via tooling.
Paper Structure (9 sections, 5 figures, 2 tables)

This paper contains 9 sections, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Regulation and Licensing for AI
  3. A Study of AI Licenses
  4. Methodology
  5. Interviews with License Adopters
  6. Analysis of Licenses and their Clauses
  7. Discussion
  8. Conclusion
  9. Tree Map of Models by License

Figures (5)

  • Figure 1: Adoption of Licenses with Behavioral Use Clauses. The number of repositories by license type on the HuggingFace model hub. As of January 2024, 41,700 RAIL licensed repositories and 3,566 LLaMA2 licensed repositories existed, both of these licenses include behavioral-use clauses.
  • Figure 2: Questions posed to our interviewees regarding the adoption of responsible AI licenses.
  • Figure 3: Applications of Models by License Type. The number of repos in each application domain that are under each license.
  • Figure 4: Interface Designs for a Responsible AI License Generator.1. Setup panel for choosing a license for "placeholderAI." On the left hand side, the type of artifact(s) specification and derivation handling can be specified by the user. 2. Customization panel for adding further use restrictions. On the left hand side, a dropdown menu allows selection of use restrictions by domain. 3. On the right hand side, ticked selections are added to the preview in a separate block at the bottom. Note that icons related to the domains of selected restrictions are added to the top row.
  • Figure 5: Model Treemap by Downloads. Treemap showing the top 10 models for each license weighed by the number of downloads